Video Segmentation Of Dynamic Multiresolution Spatiotemporal Model Based On Quaternion Wavelet Transform

How to exactly segment the object from the video sequences with complexbackground is an important problem in the field of video segmentation. At present,algorithms based on background modeling as well as based on prior training forspecific target are two main video segmentation algorithms. However, the former isvery sensitive to the degree of camera movement and the latter requires prior trainingfor specific goals. Therefore, we hope to establish the model which can segment thevideo without fixing the camera position and prior training the specific goals.In this paper, we propose a dynamic spatiotemporal saliency model based on thequaternion wavelet transform for video segmentation, which has the capability ofsegmenting the salient objects from moving background automatically. First of all, weestablish the multiresolution image pyramid with the quaternion wavelet transform.The phase can be calculated by the details extracted from the iterative approximationof pyramid. And then, we compute the motion saliency of the target object by usingthe phase disparity of the consecutive images and establish the motion magnitudemodel. Then we build the motion field model based on motion salient features. In thisway we finally establish the temporal attention model. Second, we establish a centerregion centered at the wavelet coefficients and a surrounding region, calculating thecontrast energy. Then we use Contrast Sensitivity Function proposed by Mullen toweight the center-surround contrast energy and modulate the weight of correspondingcenter region. At last, we perform the inverse quaternion wavelet transform on weightto compute the color saliency of images. Static spatial model can be established bythis way. Thirdly, we combine the two kinds of attention information together to getthe preliminary results and use Grabuct algorithm for the optimization of the resultfinally. The segmentation and comparison experimental results demonstrate thevalidity of proposed algorithm.In the past work, we optimize the preliminary results with Grabcut algorithm.Because many segmentation algorithm segments an image in pixel level, the targetand background will be segmented inaccurately on condition that color of both aresimilar. Besides, bigger size of images makes slower speed of segmentation.Therefore, we propose an segmentation algorithm based on superpixel and Graphcutframework. Our algorithm first pretreat the images to obtain many areas with good area consistency and edge descriptiveness, named superpixels. Secondly, we applyFCM to superpixels to get preliminary saliency results. At last, we optimize the resultsbased on Graphcut framework, making segmentation boundary more smooth andaccurate.